In S.Li and A.Jain,(ed).Handbook of Face Recognition.Springer-Verlag,2005
Face Databas
Ralph Gross
The Robotics Inistitute,Carnegie Mellon University
5000Forbes Avenue,Pittsburgh,PA15213
Email:{rgross}@cs.cmu.edu
Becau of its nonrigidity and complex three-dimensional(3D)structure,the appearance of a face is affected by a large number of factors including identity,face po,illumination,facial expression,age,occlusion,and facial hair.The develop-ment of algorithms robust to the variations requires databas of sufficient size that include carefully controlled variations of the factors.Furthermore,common databas are necessary to comparatively evaluate algorithms.Collecting a high quality databa is a resource-intensive task:but the availability of public face databas is important for the advancement of thefield.In this chapter we review27publicly available databas for face recognition,face detection,and facial expression analysis.
1Databas for Face Recognition
Face recognition continues to be one of the most popular rearch areas of computer vision and machine learning.Along with the development of face recognition algorithms,a comparatively large number of face databas have been collected. However,many of the databas are tailored to the specific needs of the algorithm under development.In this ction we review publicly available databas that are of demonstrated u to others in the community.At the beginning of each subction a table summarizing the key features of the databa is provided,including(where available)the number of subjects,recording conditions,image resolution,and total number of images.Table1gives an overview of the recording conditions for all databas discusd in this ction.Owing to space constraints not all databas are discusd at the same level of detail.Abbreviated descriptions of a number of mostly older databas are included in Section1.13.The scope of this ction is limited to databas containing full face imagery.Note,however,that there are databas of subface images available,such as the recently relead CASIA Iris databa[23].
1.1AR Databa
No.of subjects Image Resolution
4
4
3288
116
2
2
face
Databa No.of subjects Illumination Time
14
BANCA208++12
216
CMU Hyper5441–5
133
Equinox IR9131
9–202
瞒上欺下
Harvard RL1077–841
75
MIT1531
31
ND HID300+310/13
2++
ORL10++++
++++
U.Texas28411
爱就色色11
XM2VTS29514
16
Yale B10641
Table1:Overview of the recording conditions for all databas discusd in this ction.Cas where the exact number of conditions is not determined(either becau the underlying measurement is continuous or the condition was not controlled for during recording)are marked with“++.”
1.2BANCA Databa
No.of Subjects Image Resolution
Image quality
208
12
The BANCA multi-modal databa was collected as part of the European BANCA project,which aimed at developing and implementing a cure system with enhanced identification,authentication,and access control schemes for applications over the Internet[1].The databa was designed to test multimodal identity verification with various acquisition devices (high and low quality cameras and microphones)and under veral scenarios(controlled,degraded,and adver).Data were collected in four languages(English,French,Italian,Spanish)for52subjects each(26men and26women).Each subject was recorded during12different ssions over a period of3months.Recordings for a true client access and an informed imposter attack were taken during each ssion.For each recording the subject was instructed to speak a random12-digit number along with name,address,and date of birth(client or imposter data).Recordings took an average of20conds. Figure2shows example images for all three recording conditions.The BANCA evaluation protocol specifies training and testing ts for a number of experimental configurations,so accurate comparisons between algorithms are possible.
(1)(2)(3)(4)(5)
(6)(7)(8)(9)(10)
(11)(12)(13)
Figure 1:AR databa.The conditions are (1)neutral,(2)smile,(3)anger,(4)scream,(5)left light on,(6)right light on,(7)both lights on,(8)sun glass,(9)sun glass/left light (10)sun glass/right light,(11)scarf,(12)scarf/left light,(13)scarf/right light
1.3CAS-PEAL Databa
No.of Subjects
Image Resolution
104021377643862339–1530,900
2972–42961–266
2
1The
construction of the CAS-PEAL face databa has been supported by the China National Hi-Tech Program 2001AA114010.健康管理考试
Controlled Degraded Adver
Figure 2:Images for the three recording conditions in the BANCA databa.A high quality digital camera was ud to record the images for the controlled and adver conditions.The images of the degraded condition were taken with a low quality web cam.
subjects returned 6months later for additional recordings.Of the 99,594images in the databa,30,900images are available in the current relea.To facilitate databa distribution,the rel
ea images are stored as cropped gray-scale images of size 360×480.Figure 5shows example images of the currently distributed images.
1.4CMU Hyperspectral Face Databa
No.of Subjects
Spectral Range
454
1–5
640×480
4110.html
The CMU Hyperspectral databa,collected under the DARPA HumanID program [28],covers the visible to near-infrared range from 0.45to 1.1µm [7].Using a CMU developed imaging nsor bad on an Acousto-Optic-Tunable Filter (AOTF),the wavelength range is sampled in 10nm steps,resulting
in 65images.Acquisition of the 65images took an average of 8conds.Becau of the relative lack of nsitivity of the system (only 5–10%of light is ud),comparatively strong illumination from one to three 600W halogen lamps was ud during data collection.The lamps were placed at -45◦,0◦,and +45◦with respect to the subject.Each of the 54subjects was then imaged under four illumination conditions (three lamps individually and then combined).Subjects were recorded between one and five times over a 6–week period.Figure 6shows example images for a lection of wavelengths between 0.5and 1µm .
1.5CMU Po,Illumination,and Expression (PIE)Databa
有线宽带No.of Subjects
Image Resolution
1368
4341,368
3
夏季护肤418.html
The CMU PIE databa was collected between October and December 2000[38].It systematically samples a large number of po and illumination conditions along with a variety of facial expressions.Although only available for 2years,PIE has
Figure3:Po variation in the CAS-PEAL databa.The images were recorded using parate cameras triggered in clo succession.The cameras are each about22.5◦apart.Subjects were asked to look up,to look straight ahead,and to look down.Shown here are ven of the nine pos currently being distributed.
打印机脱机
Figure4:Illumination variation in the CAS-PEAL databa.The images were recorded with constant ambient illumination and manually triggeredfluorescent lamps.
already made an impact on algorithm development for face recognition across po[15,33]and on the evaluation of face recognition algorithms[16].So far the databa has been distributed to more than150rearch groups.
The PIE databa contains41,368images obtained from68individuals.The subjects were imaged in the CMU3D Room [21]using a t of13synchronized high-quality color cameras and21flashes.The resulting RGB color images are640×480in size.Figure7shows example images of a subject in all13pos.In addition to the po quence,each subject was recorded under four additional conditions.
1.Illumination1:A total of21flashes are individually turned on in rapid quence.The images in the illumination
1condition were captured with the room lights on,which produces more natural looking images than the cond condition.Each camera recorded24images,2with noflashes,21with oneflashfiring,and then afinal image with no flashes.Only the output of three cameras(frontal,three-quarter,and profile view)was kept.
2.Illumination2:The procedure of the illumination1condition was repeated with the room lights off.The output of all
13cameras was retained in the databa.Combining the two illumination ttings,a total of43illumination conditions were recorded.
英雄联盟英文名字
Figure5:Example relea images of the po subt of the CAS-PEAL databa.Images are gray-scale and360×480in size.
0.5µm0.6µm0.7µm
0.8µm0.9µm1µm
Figure6:Example images of the CMU Hyperspectral Face Databa.Each recording produces images for every 10nm step in the range between0.45and1.1µm.
3.Expression:The subjects were asked to display a neutral face,to smile,and to clo their eyes in order to simulate a
blink.The images of all13cameras are available in the databa.
4.Talking:Subjects counted starting at1for2conds.60frames of them talking were recorded using three cameras
(frontal,three-quarter,and profile views).
Examples of the po and illumination variation are shown in Figure8.Figure8a contains variations with the room lights on and Figure8b with the lights off.
In addition to the raw image data,a variety of miscellaneous“meta-data”were also collected to aid in calibration and other processing.
Head,camera,andflash locations:Using a theodolite,the xyz locations of the head,the13cameras,and
西安石油大学是几本the21flashes were measured.The numerical values of the locations are included in the databa and can be ud to estimate(relative) head pos and illumination directions.
Background images:At the start of each recording ssion,a background image was captured from each of the13cameras.
The images can be ud for background subtraction to help localize the face region.
Color calibration images:Although the cameras that were ud are all of the same type,there is still a large amount of variation in their photometric respons due to their manufacture and to the fact that the aperture ttings on the
